Fix a hole in the VFP thread migration. Lets define two threads.
Thread 1, we'll call 'interesting_thread' which is a thread which is
running on CPU0, using VFP (so vfp_current_hw_state[0] =
&interesting_thread->vfpstate) and gets migrated off to CPU1, where
it continues execution of VFP instructions.
Thread 2, we'll call 'new_cpu0_thread' which is the thread which takes
over on CPU0. This has also been using VFP, and last used VFP on CPU0,
but doesn't use it again.
The following code will be executed twice:
cpu = thread->cpu;
/*
* On SMP, if VFP is enabled, save the old state in
* case the thread migrates to a different CPU. The
* restoring is done lazily.
*/
if ((fpexc & FPEXC_EN) && vfp_current_hw_state[cpu]) {
vfp_save_state(vfp_current_hw_state[cpu], fpexc);
vfp_current_hw_state[cpu]->hard.cpu = cpu;
}
/*
* Thread migration, just force the reloading of the
* state on the new CPU in case the VFP registers
* contain stale data.
*/
if (thread->vfpstate.hard.cpu != cpu)
vfp_current_hw_state[cpu] = NULL;
The first execution will be on CPU0 to switch away from 'interesting_thread'.
interesting_thread->cpu will be 0.
So, vfp_current_hw_state[0] points at interesting_thread->vfpstate.
The hardware state will be saved, along with the CPU number (0) that
it was executing on.
'thread' will be 'new_cpu0_thread' with new_cpu0_thread->cpu = 0.
Also, because it was executing on CPU0, new_cpu0_thread->vfpstate.hard.cpu = 0,
and so the thread migration check is not triggered.
This means that vfp_current_hw_state[0] remains pointing at interesting_thread.
The second execution will be on CPU1 to switch _to_ 'interesting_thread'.
So, 'thread' will be 'interesting_thread' and interesting_thread->cpu now
will be 1. The previous thread executing on CPU1 is not relevant to this
so we shall ignore that.
We get to the thread migration check. Here, we discover that
interesting_thread->vfpstate.hard.cpu = 0, yet interesting_thread->cpu is
now 1, indicating thread migration. We set vfp_current_hw_state[1] to
NULL.
So, at this point vfp_current_hw_state[] contains the following:
[0] = &interesting_thread->vfpstate
[1] = NULL
Our interesting thread now executes a VFP instruction, takes a fault
which loads the state into the VFP hardware. Now, through the assembly
we now have:
[0] = &interesting_thread->vfpstate
[1] = &interesting_thread->vfpstate
CPU1 stops due to ptrace (and so saves its VFP state) using the thread
switch code above), and CPU0 calls vfp_sync_hwstate().
if (vfp_current_hw_state[cpu] == &thread->vfpstate) {
vfp_save_state(&thread->vfpstate, fpexc | FPEXC_EN);
BANG, we corrupt interesting_thread's VFP state by overwriting the
more up-to-date state saved by CPU1 with the old VFP state from CPU0.
Fix this by ensuring that we have sane semantics for the various state
describing variables:
1. vfp_current_hw_state[] points to the current owner of the context
information stored in each CPUs hardware, or NULL if that state
information is invalid.
2. thread->vfpstate.hard.cpu always contains the most recent CPU number
which the state was loaded into or NR_CPUS if no CPU owns the state.
So, for a particular CPU to be a valid owner of the VFP state for a
particular thread t, two things must be true:
vfp_current_hw_state[cpu] == &t->vfpstate && t->vfpstate.hard.cpu == cpu.
and that is valid from the moment a CPU loads the saved VFP context
into the hardware. This gives clear and consistent semantics to
interpreting these variables.
This patch also fixes thread copying, ensuring that t->vfpstate.hard.cpu
is invalidated, otherwise CPU0 may believe it was the last owner. The
hole can happen thus:
- thread1 runs on CPU2 using VFP, migrates to CPU3, exits and thread_info
freed.
- New thread allocated from a previously running thread on CPU2, reusing
memory for thread1 and copying vfp.hard.cpu.
At this point, the following are true:
new_thread1->vfpstate.hard.cpu == 2
&new_thread1->vfpstate == vfp_current_hw_state[2]
Lastly, this also addresses thread flushing in a similar way to thread
copying. Hole is:
- thread runs on CPU0, using VFP, migrates to CPU1 but does not use VFP.
- thread calls execve(), so thread flush happens, leaving
vfp_current_hw_state[0] intact. This vfpstate is memset to 0 causing
thread->vfpstate.hard.cpu = 0.
- thread migrates back to CPU0 before using VFP.
At this point, the following are true:
thread->vfpstate.hard.cpu == 0
&thread->vfpstate == vfp_current_hw_state[0]
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
Rename this branch to more accurately reflect why its taken, rather
than what the following code does. It is the only caller of this code.
This helps to clarify following changes, yet this change results in no
actual code change.
Document the VFP hardware state at the target of this branch.
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
Rename the slightly confusing 'last_VFP_context' variable to be more
descriptive of what it actually is. This variable stores a pointer
to the current owner's vfpstate structure for the context held in the
VFP hardware.
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
Directives such as .long and .word do not magically cause the
assembler location counter to become aligned in gas. As a result,
using these directives in code sections can result in misaligned
data words when building a Thumb-2 kernel (CONFIG_THUMB2_KERNEL).
This is a Bad Thing, since the ABI permits the compiler to assume
that fundamental types of word size or above are word- aligned when
accessing them from C. If the data is not really word-aligned,
this can cause impaired performance and stray alignment faults in
some circumstances.
In general, the following rules should be applied when using data
word declaration directives inside code sections:
* .quad and .double:
.align 3
* .long, .word, .single, .float:
.align (or .align 2)
* .short:
No explicit alignment required, since Thumb-2
instructions are always 2 or 4 bytes in size.
immediately after an instruction.
Reviewed-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Dave Martin <dave.martin@linaro.org>
Acked-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
vfp_put_double() takes the double value in r0,r1 not r1,r2.
Reported-by: Tarun Kanti DebBarma <tarun.kanti@ti.com>
Cc: <stable@kernel.org>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
This CPU generates synchronous VFP exceptions in a non-standard way -
the FPEXC.EX bit set but without the FPSCR.IXE bit being set like in the
VFP subarchitecture 1 or just the FPEXC.DEX bit like in VFP
subarchitecture 2. The main problem is that the faulty instruction
(which needs to be emulated in software) will be restarted several times
(normally until a context switch disables the VFP). This patch ensures
that the VFP exception is treated as synchronous.
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Cc: Nicolas Pitre <nico@cam.org>
We've observed that ARM VFP state can be corrupted during VFP exception
handling when PREEMPT is enabled. The exact conditions are difficult
to reproduce but appear to occur during VFP exception handling when a
task causes a VFP exception which is handled via VFP_bounce and is then
preempted by yet another task which in turn causes yet another VFP
exception. Since the VFP_bounce code is not preempt safe, VFP state then
becomes corrupt. In order to prevent preemption from occuring while
handling a VFP exception, this patch disables preemption while handling
VFP exceptions.
Signed-off-by: George G. Davis <gdavis@mvista.com>
Acked-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
This patch adds ptrace support for setting and getting the VFP registers
using PTRACE_SETVFPREGS and PTRACE_GETVFPREGS. The user_vfp structure
defined in asm/user.h contains 32 double registers (to cover VFPv3 and
Neon hardware) and the FPSCR register.
Cc: Paul Brook <paul@codesourcery.com>
Cc: Daniel Jacobowitz <dan@codesourcery.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
When CONFIG_PM is selected, the VFP code does not have any handler
installed to deal with either saving the VFP state of the current
task, nor does it do anything to try and restore the VFP after a
resume.
On resume, the VFP will have been reset and the co-processor access
control registers are in an indeterminate state (very probably the
CP10 and CP11 the VFP uses will have been disabled by the ARM core
reset). When this happens, resume will break as soon as it tries to
unfreeze the tasks and restart scheduling.
Add a sys device to allow us to hook the suspend call to save the
current thread state if the thread is using VFP and a resume hook
which restores the CP10/CP11 access and ensures the VFP is disabled
so that the lazy swapping will take place on next access.
Signed-off-by: Ben Dooks <ben-linux@fluff.org>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
On ARMv7, conditional undefined instructions may generate exceptions
even if the condition is not met. The vfphw.S contains the FPINST and
FPINST2 access instructions which may not be present on processors with
synchronous VFP exceptions.
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
This declaration specifies the "function" type and size for various
assembly functions, mainly needed for generating the correct branch
instructions in Thumb-2.
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
This patch adds the support for VFPv3 (the kernel currently supports
VFPv2). The main difference is 32 double registers (compared to 16).
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
This patch allows the VFP support code to run correctly on CPUs
compatible with the common VFP subarchitecture specification (Appendix
B in the ARM ARM v7-A and v7-R edition). It implements support for VFP
subarchitecture 2 while being backwards compatible with
subarchitecture 1.
On VFP subarchitecture 1, the arithmetic exceptions are asynchronous
(or imprecise as described in the old ARM ARM) unless the FPSCR.IXE
bit is 1. The exceptional instructions can be read from FPINST and
FPINST2 registers. With VFP subarchitecture 2, the arithmetic
exceptions can also be synchronous and marked by the FPEXC.DEX bit
(the FPEXC.EX bit is cleared). CPUs implementing the synchronous
arithmetic exceptions don't have the FPINST and FPINST2 registers and
accessing them would trigger and undefined exception.
Note that FPEXC.EX bit has an additional meaning on subarchitecture 1
- if it isn't set, there is no additional information in FPINST and
FPINST2 that needs to be saved at context switch or when lazy-loading
the VFP state of a different thread.
The patch also removes the clearing of the cumulative exception flags in
FPSCR when additional exceptions were raised. It is up to the user
application to clear these bits.
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
The current lazy saving of the VFP registers is no longer possible
with thread migration on SMP. This patch implements a per-CPU
vfp-state pointer and the saving of the VFP registers at every context
switch. The registers restoring is still performed in a lazy way.
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
Patch from Daniel Jacobowitz
vfp_put_double didn't work in a CONFIG_AEABI kernel. By swapping
the arguments, we arrange for them to be in the same place regardless
of ABI. I made the same change to vfp_put_float for consistency.
Signed-off-by: Daniel Jacobowitz <dan@codesourcery.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
Since we pass flags to the compiler to control code generation based
on the least capable selected CPU, if we want to include VFP support,
we must tweak the assembler flags to allow the VFP instructions.
Moreover, we must not use the mrrc/mcrr versions since these will not
be recognised by the assembler.
We do not convert all instructions to the VFP-equivalent (yet) since
binutils appears to barf on "fmrx rn, fpinst" and doesn't provide any
other way (other than using the mrc equivalent) to encode this
instruction - which is rather a problem when you have a VFP
implementation which requires these instructions.
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
Patch from Catalin Marinas
This patch changes the double registers numbering to 0-15 from even 0-30,
in preparation for future VFP extensions. It also fixes the VFP_REG_ZERO
bug (value 16 actually represents the 8th double register with the original
numbering).
The original mcrr/mrrc on CP10 were generating FMRRS/FMSRR instead of
FMRRD/FMDRR. The patch changes to CP11 for the correct instructions.
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
Patch from Catalin Marinas
The current VFP code corrupts the VFP registers (including the control
ones) if more than one floating point application is executed at the same
time. This patch fixes the updating of the load/store base addresses for
the VFP registers.
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
Initial git repository build. I'm not bothering with the full history,
even though we have it. We can create a separate "historical" git
archive of that later if we want to, and in the meantime it's about
3.2GB when imported into git - space that would just make the early
git days unnecessarily complicated, when we don't have a lot of good
infrastructure for it.
Let it rip!
Russell King